Re: singularity and exi-GIMPS team

dan (
Thu, 12 Aug 1999 15:39:35 -0400

"Eliezer S. Yudkowsky" wrote:
> Spike, you're not in my kill file.
> Yes, distributed computing advances the Singularity. I think the people
> at have a bunch of old Pentiums networked together. I'm
> not sure an ExI network would help, unless there was new software coming
> out as a result. I don't know of anyone here who's working on
> AI/neural-net stuff that runs distributed, someone who could benefit directly.
> --
Actually, there are some interesting algorithms that can use a very loosely coupled system. Many parallel algorithms need high bandwidth and low latency, but others don't. The current Key finders and SETI search are at the extreme end of a spectrum, and can make effective use of a large number of intermittently-connected processors with extremely low bandwidth. At the other end are algorithms that are bandwidth-constrained even when the processors are using processor busses or shared memory for interprocessor communication.

The trick is to find a useful algorithm that uses the environment that is now emerging. Many of us will soon have full-time connections (cable or ADSL) with bandwidth in the 100Kbps range or better. Just as the advent of lots of intermittently-connected Pentium-class computers gave rise to the Key searchers, etc. cable modems and ADSL will enable a new class of distributed algorithms.

It looks to me like genetic algorithms may be very effective in such an environment. It appears that you can send a small "population" of solutions off to a computer and have that computer "breed" an improved population, and then "cross-breed" the resulting population back into the larger "gene pool." With a simple fitness function, the intergeneration time is low and the bandwidth high, but a more complex fitness function can be employed to gain better results if you have enough computers. This lowers the bandwidth per computer, and I think we can find some neat singularity-advancing genetic algorithms.

I don't know if this can be used an the first step of the "singularity bootstrap" we discussed three years ago ( ) But maybe it could be. If we can create a fitness function for optimizing fitness functions, and a fitness function for evaluating how well a program works, we will be on our way to optimizing the web.